4 ms·

drop horizontal layers architectures and focus on vertical-slices. This alone reduces the needed context window per feature by an order of magnitude.